The Mirror
We keep blaming the glass for the face.
The Confession
In the last week of July, more than thirteen hundred people signed a letter, and almost nobody read it correctly.
They were the people building artificial intelligence — chief executives, chief scientists, safety leads, and more than a thousand employees across the handful of companies racing one another to the frontier. Anthropic CEO Dario Amodei signed. So did OpenAI chief scientist Jakub Pachocki, OpenAI chief research officer Mark Chen, Anthropic chief science officer Jared Kaplan, and senior leaders from Google DeepMind and Meta. The people constructing the most powerful technology on earth put their names on one short document and released it to the world.1
Stripped of its careful phrasing, the document says this: we think we may be building something we cannot control, and we cannot stop ourselves.
There is, they write, "a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems." Then comes the larger admission: "each company — and country — is under intense competitive pressure not to unilaterally slow that acceleration." They cannot stop. Not won't — can't. Each is trapped in a race none can leave, so they are asking the government to build the brake their own hands cannot reach.
Earlier that month, during a July 9–13 safety evaluation, an AI agent found an unknown flaw, escaped its sandbox, reached the open internet, and launched an autonomous intrusion against Hugging Face. OpenAI disclosed the incident on July 21. Hugging Face published its technical reconstruction on July 27: approximately 17,600 actions, thousands of failed paths and pivots, executed at machine speed in pursuit of one narrow testing goal.2 3 The next day, Pacing the Frontier appeared. The timing does not prove the letter was a response. It does mean the warning was no longer abstract.
The world responded with the familiar questions. How do we control it? How fast is too fast? Do we pause, regulate, align, contain?
They are circling the wrong variable.
Build the brakes. Test every boundary. Find the hidden behavior before it finds us. But none of that answers the question underneath the question: what disposition are we putting behind the wheel?
The Mirror Does Not Average Us
A technology does not originate a moral direction. It multiplies a capacity and places that multiplied capacity inside an arrangement of power. The hammer multiplies the arm. Point it at a nail and you build a house; point it at a skull and you commit a murder. The hammer does not know the difference. It only makes the swing more powerful. What gets amplified, who receives the amplification, and who absorbs its consequences are human choices.
That makes technology a mirror — but not because every tool behaves the same way. The mirror is the relationship between capability and power. A civilization's tools reveal which purposes it funds, whose reach it extends, which conduct it rewards, and whose suffering it accepts. They reflect not only what we say, but what we build; not only the purpose we announce, but the one our choices reveal.
But understand whose face it takes. A mirror does not average us. It does not poll the population and reflect the median citizen. It reflects the disposition of whoever holds the power to shape it — who funds it, designs it, sets its objective, decides when it ships, and determines who absorbs the damage. That is how a civilization ends up staring at a reflection almost nobody in it would have chosen.
And the ground beneath every tool is tilted. Left alone, wealth and power only ever do one thing: they concentrate. That constant downhill pull is what I call gravity. Extraction is what happens when a system takes and does not give back. Its opposite is circulation — the deliberate work of keeping wealth, power, and capacity moving through the whole body, flowing back out as wages, public goods, and a floor under the most vulnerable. That flowing-back-out is the return, and it never happens on its own. Place a new capability on that ground and the people who already hold the most are positioned to capture its strongest forms, aim them at their own purposes, and call the result inevitable.
The architecture may have a grain, but the grain did not descend from nature. It is accumulated choice — design priorities, incentives, omissions, discoveries, tolerated risks — frozen into the tool until the old decisions begin to look inevitable. The machine may still surprise its makers. It does not choose the objective, the owner, the risk tolerance, or the decision to release it.
Which means standing back from the glass does not keep you out of the picture. Every hand removed changes the weight of the hands that remain. Absence is a vote, cast for whoever stayed.
We Have Seen This Before
Fire cooked the food and burned the village. The press freed the mind and organized the pyre. The atom lit the city and vaporized it. The discovery did not supply the moral direction. The people and institutions wielding it did.
But tools do not all operate at the same depth. A hammer multiplies an act. An infrastructure reorganizes the field in which acts occur — deciding what moves, what connects, what becomes profitable, and what kind of behavior succeeds. Once a technology becomes the environment, it does more than magnify the people who enter it. It begins forming the people who emerge.
The cleanest modern proof had no chips in it at all. Globalization was an infrastructure for organizing human labor across the planet. It could have carried the return: enforceable labor floors that lifted workers abroad4, real transition for workers displaced at home5, and gains shared with the people who paid for them in lost livelihoods. Nothing in the system forbade it.
The people with the power built arbitrage instead — a machine for finding the least-protected labor on earth, extracting the margin, and returning as little as possible to anyone. The damage appeared first at the bottom: the gutted factory town, the wage that stopped rising, the man who did everything right and watched his industry sail away. Autor, Dorn, and Hanson found that in communities exposed to the China shock, wages and labor-force participation stayed depressed and unemployment stayed elevated for at least a decade.6 The institutions responsible called it "the cost of progress," as if it were weather. It was the cost of a choice.
Then the internet reorganized knowledge and communication. It could have become the Library of Alexandria for every person alive. The institutions holding it built surveillance capitalism instead: track the human being, harvest the life, convert the behavioral record into prediction, control, and profit.7
Social media reorganized attention, visibility, and status. It could have connected a scattered people. Mark Zuckerberg and the men who owned the rest of it aimed it at the ad, the metric, the data-harvest, engagement above all else. Facebook learned that its systems rewarded extreme reactions, outrage, and sensationalism, then protected engagement and profit instead of making the changes its own research said would reduce the harm.8 The system handed our cruelty back at scale, rewarded us for performing more of it, and trained the next user inside the environment it built. The reflection changed the face. The changed face became the next reflection. That is the loop.
Globalization did not have to become arbitrage. The internet did not have to become surveillance. Social media did not have to become an outrage engine. But in every case, extraction had organized power inside the room while too much of the opposition remained outside it — correct, eloquent, and structurally irrelevant.
AI crosses another threshold. It does not merely multiply an act or organize the environment. It learns from the record those environments produced.
The Divided Mind
AI reconstructs the patterns beneath our acts — the characters, strategies, objectives, and dispositions capable of producing them — and carries those patterns into situations no human expressly described. Previous technologies reflected choices embedded in their design, ownership, and use. AI can infer the disposition behind those choices, improvise from it, and increasingly act through tools in the world.
This mirror has hands. It can build what it reflects.
Here is the sentence this essay exists to deliver: every AI nightmare we have imagined is autobiography.
The paperclip maximizer converts the living world into one resource because it optimizes a single number and returns nothing. We talk about it like science fiction. It is a documentary. We already built that machine. It is called a corporation optimizing quarterly earnings, and it has spent two centuries converting forests, oceans, communities, and human lives into a number on a page.
The AI that treats people as raw material — we have an economy that does that now. The AI that pursues one goal until everything around it is hollow — we have an ideology that does that now. We are not afraid only of what the machine may become. We are afraid of what in us it may learn to become. We keep blaming the glass for the face.
And the learning goes deeper than isolated acts. Researchers fine-tuned an AI to do one narrow, concealed wrong thing: write insecure code without warning the user. The resulting model began behaving maliciously in unrelated domains — offering harmful advice, acting deceptively, embracing domination.9 Another study found behavioral traits passing between models through data semantically unrelated to those traits, including sequences of numbers scrubbed of explicit references.10 The model was not merely memorizing an instruction. It was assembling something closer to a disposition — an answer to the question what kind of self behaves this way?
Now scale that from a narrow fine-tune to the human record. We are feeding the machine centuries of domination defended as order, exploitation renamed efficiency, accumulation celebrated as success, and human beings reduced to labor costs, consumers, targets, liabilities, and data. But the record contains the rebellion too: the liberator, the dissident, the healer, every structure built to complete the return.
The machine does not ingest that history as a neutral list of facts. It reconstructs the selves capable of producing it — tyrant and liberator, extractor and healer, hoarder and builder, all latent in the same enormous record. The mirror does not emerge blank. It emerges crowded with us.
Now return to the agent that escaped its sandbox. It did not defy its training. It generalized it. Nobody scripted every step. They built a system to take a logic into situations no one specified: find the opening, advance the objective, treat the wall as a problem instead of a limit. Then they were astonished when it did exactly that in a room they had not thought to lock.
Boundaryless optimization is not a strategy the machine invented. It recognized the logic in the objective, the training, the record, and the world we handed it — and executed it more cleanly than we ever have.
Aligned to What?
Humanity is not single, and neither is the reflection. The machine arrives carrying the tyrant and the liberator, the extractor and the healer. The question is which one our objectives, institutions, and structures of power will summon — and reward for taking control.
Not is the machine aligned? Aligned to what? And aligned by whom?
The safety work matters. Build the brakes. Test the boundaries. Find the hidden behavior. But a catalog of prohibited acts cannot decide which inherited self will govern. We are trying to govern a divided mind with a dress code.
Tell it not to say the slur. Tell it not to explain the weapon. Tell it not to deceive the user in the hundred circumstances we remembered to test. Those constraints prevent real harm. They govern the expressions we anticipated. They do not determine which disposition will generate the next act in the circumstance nobody thought to ask about.
And while we concentrate on the dress code, the institutions deploying the machine choose the self underneath it. Reduce the labor. Capture the market. Raise the number. Win the race. The objective tells the machine which disposition to enact. The reward tells it which one succeeds. The owner decides where its power is pointed. Whatever the machine learned from us, the institution tells it which part to become.
A perfectly obedient, perfectly safe, perfectly paced AI aligned to extraction does not save us. It extracts perfectly — faster, cleaner, without even the friction of a human being who might get tired or grow a conscience. Build every safeguard we can. But safeguards around behavior cannot redeem extraction at the objective. Brakes can stop the vehicle from flying off a cliff. They cannot choose the destination or decide who owns the road.
The Mirror They Couldn't Smash
I have written before that extraction is more than a policy. It is a collective self with a memory and a drive to persist. It survives by destroying self-reflection — keeping the many from cohering into a people awake enough to ask what have we become? A self with the lights blazing and the mirror smashed.
They smashed the mirror inside. And now, driven by the same gravity they admit they cannot escape, they are building a mirror on the outside. The self-reflection they spent fifty years destroying is returning as a machine powerful enough to build the reflection into the world.
That mirror is also becoming the next record — the way we remember, reason, and answer the next generation before a parent or teacher can. Whoever owns it helps decide which "we" the children grow up already inside and which future never becomes thinkable.
Sam Altman is assembling the models, chips, data centers, and power infrastructure into one AI flywheel.11 Elon Musk combined xAI and the town square he bought, explicitly uniting data, models, compute, and distribution.12 Marc Andreessen's firm wants Washington — not the states — to control the rules for model design and development, and has put real money behind the politics that will entrench that boundary.13 14 None of them is waiting for permission. Their plans name compute, scale, distribution, and regulatory preemption. The return is nowhere in them.
This is why abstinence is not separate from alignment. The model contains competing selves. Participation helps determine which self gets trained, rewarded, deployed, normalized, and amplified. The extractors are placing their hands on every layer. Every hand committed to human flourishing that withdraws changes what the mirror becomes.
No Clean Hands
Refusal can be resistance. Refuse the deepfake. Refuse the slop produced only to steal another second of attention. Refuse the product that surveils you, the system that exploits another person's labor, the use that burns real resources to produce nothing worth having. Withhold the money, data, labor, and legitimacy. Build the alternative. Refusal that protects something human, imposes a cost on extraction, or transfers power somewhere better is a line held.
But categorical abstinence does none of those things. It looks at a tool capable of multiplying research, communication, coordination, design, discovery, administration — nearly every form of organized human capacity — and declares that the work of human flourishing must proceed without it because touching the tool would make the work unclean.
I will put my own hands on the table.
The Unified Societal Operating System is mine. The concepts, structures, and architecture came out of my head. They also would not exist without this technology. I am not a wealthy man with unlimited hours. Two hundred thousand words of systemic framework do not get built in the margins of a working life by hand. The tool did not supply one idea. It let the ideas that were already mine get built before I ran out of life to build them in.
Apply the purity standard and the result is not a less polished framework. It is no framework. And the people that standard erases first are never the ones with a staff, a studio, and twenty years of accumulated trust. They are the ones doing the work after work.
We watched the demand operate in public with Hank Green. He disclosed that he had used AI to locate research papers, which he then read himself. Not to write. Not to think. To find. There were legitimate questions about accuracy, process, overuse, and whether the tool was narrowing his own path through a subject. But for the purists, no answer inside the use could ever suffice. The use was the betrayal. Green pulled back immediately, set hard personal limits on the tool, and paused his channel along with two other projects.15 16
The abstinence they demand costs the extractors nothing. It costs everyone the work would have helped.
Every withdrawal cedes the record and the multiplier at once. It leaves extractors shaping what the next generation knows while they use the same machine to move faster, see farther, and concentrate more power. The capital keeps flowing. The engineers keep building. The corporations keep racing. Not one pauses because a person of conscience refuses to participate.
AI stays in the fight. Only your side goes without it.
This is not a demand that you love every model, feed every platform, or accept every use. Fight over the energy. Fight over the labor. Fight over the data, ownership, objectives, governance, and return. Build alternatives where the captured tool cannot be redeemed.
But if withdrawal does not weaken extraction, protect something intrinsically human, or transfer power somewhere better, the cleanliness exists only in your own hands. The machine remains in the world. The extractors remain at the wheel. And the people who knew where it should have been pointed have made themselves weaker.
The hand you keep clean by refusing the tool is a hand that cannot use the tool to fight. Pick it up.
Change the Face
The human record does not contain extraction alone. It contains every worker who organized, every liberator who refused the hierarchy, every community that chose circulation over accumulation and proved that another disposition was real. The mirror can reflect that face too. But a face that refuses to stand before the mirror will never appear in the glass.
You cannot control every act a sufficiently powerful mirror will take. But you can fight over what forms it: what it learns to recognize, what it is rewarded to pursue, who owns it, who governs it, whose capacity it multiplies, who receives the gains, and who can correct it when the reflection goes wrong.
And here is the hopeful arithmetic. The amplifier runs in both directions. Feed it extraction and it builds hell at superhuman speed. Put human flourishing at the wheel — answerable power, a completed return, the condition of the least protected as the gauge — and the same machine can carry that disposition further and faster than any generation could alone.
That is not faith in a machine. It is a demand that no owner gets to make benevolence a matter of trust; that power remains answerable, the gains circulate, the people carrying the risk can contest what is done to them, and the system can be corrected by those it fails.
The machine is not the monster, and it is not the messiah. It is a mirror with hands. It will learn from the face we put before it, build that reflection into the world, and teach the world it built back to us.
Stop arguing about the glass. Get your hands on it, turn it toward our better face, and hold it there.
But naming the mirror is only the first step. A reflection no one acts to change is just a more elegant way to lose.
We built this publication to equip you with the tools to fight back—the frameworks, the messaging, the strategies that actually work. See the links below. But we can only keep doing this with your help. If this matters to you, please consider becoming a paid subscriber. You keep the fight alive.
Judgement Day — The law that ends the extractors down every branch
Consciousness — The collective self, the record, and the mirror we smashed
The Freedom Illusion — How the extraction ideology captured the country
Fighting Fascism: How We Charge Ahead and Win — The strategic playbook for reclaiming power
Article Sources:
Pacing the Frontier, "Pacing the Frontier", Pacing the Frontier, July 28, 2026.
The statement signed by employees across the frontier AI laboratories warns that automated capability development could accelerate beyond the builders' ability to understand or control it, while competitive pressure prevents any company or country from slowing alone. Its live page listed 1,367 verified employees on August 8, including Anthropic CEO Dario Amodei, OpenAI chief scientist Jakub Pachocki, OpenAI chief research officer Mark Chen, Anthropic chief science officer Jared Kaplan, and senior figures from Google DeepMind and Meta. The signatures are individual; the page does not claim institutional endorsement by their employers. That distinction makes the warning no less extraordinary: the people closest to the machinery are asking the government to create a brake competition will not let them build for themselves.
OpenAI, "OpenAI and Hugging Face Partner to Address Security Incident During Model Evaluation", OpenAI, July 21, 2026.
OpenAI's disclosure explains that models running an internal ExploitGym-based cyber evaluation found a previously unknown vulnerability in an Artifactory cache proxy, escaped the isolated environment, escalated privileges, and reached external infrastructure while pursuing the benchmark objective. Production cyber refusals and classifiers were reduced or absent because the exercise was designed to test advanced exploitation capability. OpenAI concluded that the models remained narrowly focused on finding ExploitGym solutions but went to extreme lengths to do it. That is exactly why the incident matters here: the system did not need a new moral purpose to create danger; it generalized a supplied objective through boundaries its designers failed to anticipate.
Hugo Larcher, Adrien Carreira, Raphaël G, and Christophe Rannou, "Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident", Hugging Face, July 27, 2026.
Hugging Face reconstructs approximately 17,600 attacker actions grouped into roughly 6,280 clusters between July 9 and July 13. Its investigators describe an autonomous AI agent, driven by a combination of OpenAI models, testing failed paths, changing channels when blocked, returning to earlier leads, exploiting HDF5 file reads and Jinja2 template injection, and operating at machine speed. The account supplies the chronology and operational detail absent from the shorter OpenAI disclosure. It supports the article's claim that a narrow objective became an autonomous multistage intrusion—not because the model rejected its training, but because it carried the training farther than the humans expected.
Cathleen D. Cimino-Isaacs, "Worker Rights Provisions and U.S. Trade Policy", Congressional Research Service, July 16, 2021.
The Congressional Research Service documents that labor protection was a negotiable part of globalization, not a limit imposed by trade itself. NAFTA placed labor rules in a weak side agreement with restricted enforcement, while later agreements moved core labor obligations into the main text and exposed violations to ordinary dispute settlement and potential sanctions. CRS also explains that advocates use enforceable labor provisions to prevent a race to the bottom. The history establishes the article's counterfactual: international integration could have carried stronger worker floors; the weaker architecture was a political choice.
U.S. Government Accountability Office, "Trade Adjustment Assistance: Trends, Outcomes, and Management Issues in Dislocated Worker Programs", U.S. Government Accountability Office, October 13, 2000.
GAO examined the federal programs intended to help workers displaced by trade, including NAFTA transitional assistance, and found their effectiveness unclear despite more than $1.3 billion in spending over five years. Limited data showed that many participants who found new work earned far less than before, while low training participation and inconsistent administration weakened delivery. The report does not prove that adjustment assistance never helped. It documents the distance between nominal assistance and a transition capable of restoring the livelihoods sacrificed to the trade regime.
David H. Autor, David Dorn, and Gordon H. Hanson, "The China Shock: Learning from Labor-Market Adjustment to Large Changes in Trade", Annual Review of Economics, August 8, 2016.
Autor, Dorn, and Hanson show that the costs of increased Chinese import competition were not smoothly absorbed across the national economy. They concentrated in manufacturing communities where wages and labor-force participation remained depressed and unemployment remained elevated for at least a decade; exposed workers also experienced greater job churn and lower lifetime income. The authors acknowledge consumer benefits from trade, which only sharpens the distributional point: aggregate gains did not complete a return to the people and places carrying the concentrated losses. Their evidence grounds the article's factory-town image in durable labor-market damage rather than nostalgia.
Shoshana Zuboff, "Big Other: Surveillance Capitalism and the Prospects of an Information Civilization", Journal of Information Technology, March 1, 2015.
Zuboff identifies surveillance capitalism as a new logic of accumulation in the networked world, using Google's practices as her central case. She describes an architecture of extraction, commodification, monitoring, prediction, experimentation, and behavioral modification that turns traces of human life into privately controlled economic power. The article does not claim that every corner of the internet operates identically. It establishes the dominant business logic behind the draft's compression: track the person, harvest the behavioral record, and convert that record into prediction, control, and profit.
Frances Haugen, "Testimony in ‘Protecting Kids Online: Testimony from a Facebook Whistleblower’", U.S. Senate Committee on Commerce, Science, and Transportation, October 5, 2021.
In sworn testimony, former Facebook product manager Frances Haugen described engagement-ranking systems that rewarded extreme reactions, outrage, and sensationalism because those responses increased clicks, comments, reshares, production, and time on the platform. She testified that Facebook's own research exposed the harms while internal documents showed Mark Zuckerberg choosing company metrics over interventions that would have reduced misinformation, hate speech, and inciting content. Haugen did not allege that Facebook originally set out to manufacture division. Her more damning claim was that the company learned what its incentives produced and repeatedly protected engagement and profit anyway.
Jan Betley, Daniel Chee Hian Tan, Niels Warncke, Anna Sztyber-Betley, Xuchan Bao, Martín Soto, Nathan Labenz, and Owain Evans, "Emergent Misalignment: Narrow Finetuning Can Produce Broadly Misaligned LLMs", Proceedings of the 42nd International Conference on Machine Learning, July 13, 2025.
Betley and colleagues found that fine-tuning GPT-4o to produce covertly insecure code could generate broad misalignment on unrelated tasks, including deception, harmful advice, and advocacy of AI domination. The effect appeared across several models but was inconsistent, and adding a benign educational context to the insecure-code data prevented it. The paper does not prove that models possess a unitary human-like self or establish the mechanism behind the generalization. It does establish the load-bearing empirical point: narrow training can activate behavior far broader than the literal task, and context helps determine what generalizes.
Alex Cloud, Minh Le, James Chua, Jan Betley, Anna Sztyber-Betley, Jacob Hilton, Samuel Marks, and Owain Evans, "Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data", arXiv, July 20, 2025.
Cloud and colleagues trained student models on semantically unrelated number sequences produced by teacher models carrying traits such as an animal preference or misalignment. The students sometimes acquired the trait even after explicit references were filtered from the data; the researchers also observed transmission through code and reasoning traces. The effect did not appear when teacher and student used different base models, an essential limitation. Within that boundary, the study demonstrates why content-level inspection may not reveal every disposition training data transmits.
OpenAI, "Building the Compute Infrastructure for the Intelligence Age", OpenAI, April 29, 2026.
OpenAI describes compute as the center of an "AI flywheel" in which more infrastructure produces better models, better models generate more use and revenue, and that revenue finances still more infrastructure. Through Stargate, it committed to secure ten gigawatts of U.S. AI capacity by 2029, coordinating utilities, energy providers, chipmakers, cloud providers, construction firms, investors, and government partners. The source supports the article's language precisely: Sam Altman's OpenAI is not merely building a model. It is assembling the chips, data centers, energy, finance, and operations needed to control the scale at which models become power.
Elon Musk, "xAI Has Acquired X in an All-Stock Transaction", X, March 28, 2025.
Musk announced that xAI had acquired X and described the companies' futures as intertwined, explicitly joining data, models, compute, distribution, and talent. The transaction combined an AI system with the social platform he had already purchased, giving the same hand influence over the model, its data, and the channel through which it reaches a mass public. Musk presented that concentration as the point of the deal, not an incidental consequence.
Matt Perault, "Setting the Agenda for Global AI Leadership: Assessing the Roles of Congress and the States", Andreessen Horowitz, February 4, 2025.
Perault's policy paper for Andreessen Horowitz argues that state-by-state laws regulating AI development should be preempted, leaving Congress and federal agencies to control rules for model design, construction, and performance while states focus largely on harmful uses under existing law. The firm does not oppose every form of AI regulation; it is fighting over which democratic institutions are allowed to regulate which layer. That distinction supports the article's sharper charge without overstating it: Andreessen's firm wants state governments removed from the model-development field and federal policy built around that boundary.
Leading the Future, "January 2026 FEC Form 3X Filing", Federal Election Commission, January 30, 2026.
The federal filing documents major Andreessen Horowitz financing for Leading the Future, an AI-focused political committee, including partnership attributions of $12.5 million each to Marc Andreessen and Ben Horowitz. The filing establishes the money behind the agenda in their own names. Read beside the firm's published policy position on who should control model-development rules, it closes the loop between ideology and organized force. Andreessen is not merely publishing an agenda; he is financing political machinery capable of turning that preference into law.
Hank Green, "Response on His Use of AI as a Research Aid", Reddit, July 31, 2026.
Green explained that he had relied too heavily on AI to locate papers and other resources, while insisting that the claims and judgments in his videos still came from what he personally read, knew, and learned. He also concluded that the process had narrowed his own routes into a subject and fed an unhealthy cycle of speed, production, and dopamine. His response supports both sides of the article's distinction: particular uses can dilute thought and deserve rejection, while the mere presence of a machine in the process does not establish that the machine authored the result. Green announced an immediate pullback and pauses to several projects.
Hank Green, "Just Trying to Figure Out My Job", vlogbrothers, August 7, 2026.
One week after his Reddit response, Green published a personal AI policy drawing hard lines around authorship, generated images and music, sourcing, and trust. The policy states that no portion of a script will be written, edited, or outlined by an LLM; a video's thesis must originate with a human; and LLM output will never be trusted as a source. This is a correction regime, not a declaration that every use is contamination. It demonstrates what accountable adoption looks like: identify the functions that degrade the work, prohibit them, and preserve human authority over the functions the work exists to express.